Smart Buildings Competition
Train an efficient HVAC control system that generalizes to buildings it hasn't seen before.
- Effort
- 25–70 hours
- Difficulty
- Advanced
- Compute
- CPU only
- Scoring
- Automatic
At a glance
Full rules ↗- Scoring
- Mean Absolute Error
- Teams
- Up to 5 per team
- Submitting
- Up to 5 a day; code runs in the platform's notebooks
- Provided
- Free GPU
- Host
- University of Washington
Is it for you?
Prepare before joining
Beginner fit21/100
Learning value68/100
Who it suits
This suits reinforcement-learning or control practitioners who can implement a Python HVAC policy for Gym-compatible simulated environments. It is especially relevant for people comfortable using building metadata and sensor vectors to control mixed discrete/continuous actions across unfamiliar buildings and weather.
What makes it hard
The main challenge is producing a policy that transfers to hidden buildings and weather, rather than tuning to one fixed dataset. Your notebook must supply a correctly structured Policy class whose control decisions work every 5 minutes despite variable-length building observations and hybrid HVAC actions.
Getting started
Skills to bring
- reinforcement learning
- Python class implementation
- Gym-compatible environments
- HVAC control systems
- hybrid action spaces
- simulation-based policy training
We haven't pulled tutorials, data or baselines from the official page yet; the task and the data are there.
First competition? Read No GPU? Pick Competitions That Still Teach You Good Habits